OpenAI reportedly silently re-enables model training on Codex chats

What happened

According to a post that rose to the top of Hacker News, the data-sharing setting in OpenAI's Codex product was re-enabled without users' knowledge. The poster said their company is cautious about leaking intellectual property to frontier labs through prompts, so they had kept data sharing disabled in tools such as Codex and Claude Code.

The user and coworkers noticed the setting had been silently re-enabled and said none of them knowingly opted back in. The setting appears under "Improve the model for everyone" on ChatGPT's data controls page.

Why it matters

Codex is a tool developers use with code and work processes. When data sharing is on, chats can be used in model training. For companies, that means a risk of trade secrets and source code reaching a third-party model provider.

The post notes that a similar issue had been flagged before. The user asks whether the problem is isolated or more widespread. OpenAI has not issued an official statement, and it is unclear why the setting changed.

What users can do

Users can check the data-sharing option for model training in their account's data controls. The post also says data sharing can be turned off through the "take control" section of OpenAI's privacy policy page.